Skip to content

Laguna S 2.1 Heavy qualification and promotion

Date: 2026-07-26

Host: Primary Node, NVIDIA RTX PRO 6000 Blackwell 96 GB

Decision: Promote Laguna S 2.1 NVFP4 as the managed Heavy serve with thinking disabled; retain GPT-OSS Puzzle 88B as the declared rollback.

Exact tested configuration

Field Qualified value
Checkpoint poolside/Laguna-S-2.1-NVFP4
Revision 07614121b31898586430f189d27a25a0be310843
Served name laguna-s-2.1-nvfp4
Engine vLLM 0.23.1rc1.dev1327+gf25953cc5
Image vllm/vllm-openai:nightly-f25953cc59f9b4ba9b04b16228d2b86dcfbcbdb1
Context 262,144 tokens
Managed endpoint http://127.0.0.1:30002/v1
Thinking control chat_template_kwargs.enable_thinking=false

The checkpoint identity, engine, context, and request-level thinking control are recorded in the repeated quality artifact. The source recipe and managed manifest pin the same model revision and engine image.

Why thinking is disabled

The upstream model card documents thinking as enabled by default and documents request-level control through chat_template_kwargs. A July 23 upstream discussion likewise confirms that the template toggle is the supported way to disable reasoning. The locally cached template defaults enable_thinking to true and preserve_thinking to false.

The first thinking-enabled smoke check consumed the complete 4,352-token completion allowance, returned finish_reason=length, emitted 18,840 reasoning characters, and produced no visible answer after 45.9 seconds. The immediate full rerun passed, including a 2,090-character reasoning trace in its smoke check. These two outcomes establish an intermittent exhaustion risk rather than a deterministic parser failure. The failing operator observation and passing thinking-enabled rerun are both retained.

Thinking-disabled preflight passed smoke, structured JSON, a 4K needle, and 2/2 parallel tool calls with no reasoning field. The production router and promotion plan therefore force enable_thinking=false; callers do not inherit the checkpoint's default.

Local qualification

The repeated protocol-v3 quality run used three attempts per deterministic check, a required pass rate of 1.0, 256 visible-answer tokens, and zero reasoning headroom. It recorded no failures:

Gate Result
Context retrieval Passed at 32K, 128K, and 240K targets
Actual prompt tokens 27,799; 148,899; 243,641
Context TTFT 2.26 s; 21.15 s; 50.64 s
OpenAI tool call 3/3
Multi-turn recall 3/3
Unified diff 3/3
Timeout triage 3/3

The quality artifact contains the complete visible answers, deterministic checks, budgets, usage, and latencies. A second thinking-disabled preflight and the final promotion functional gate also passed; the latter exercised a 120K requested needle and 10/10 tools.

Short-output capacity completed 10/10 requests at concurrency one and 40/40 at concurrency eight with independent 8,192-token prompts:

Concurrency TTFT p50 / p95 E2E p50 / p95 Aggregate output
1 0.07 / 0.55 s 0.57 / 0.90 s 75.46 tok/s
8 3.44 / 4.37 s 3.87 / 4.87 s 83.24 tok/s

These are short-output batch-capacity measurements, not controlled long-decode throughput. See the c1 and c8 artifacts.

Promotion and rollback

The human-approved guarded promotion:

  1. loaded the exact managed Laguna Heavy service on port 30002;
  2. passed the thinking-disabled promotion functional gate;
  3. promoted the matching router config and profile;
  4. verified router gateway HTTP 200 and exact served identity; and
  5. left the Heavy reservation in the admitting state.

GPT-OSS Puzzle 88B remains the only declared Heavy rollback. Its container is not kept resident because it competes for the same GPU, but its pinned image, checkpoint cache, manifest entry, router configuration, and prior qualification remain available through the managed rollback path.

Scope and caveats

  • This proves the exact Laguna S configuration on one RTX PRO 6000; it is not a general cross-model intelligence ranking.
  • The thinking-enabled failure is intermittent. The disabled-thinking profile is the qualified production contract, not a claim that reasoning can never work.
  • The capacity artifacts use short completions, so their aggregate output rates must not be compared with controlled long-generation numbers.
  • Promotion was human-authorized after the local gates; benchmark completion alone did not change production.